Accelerating Long-Term Xenon Dynamics Prediction: A Reduced-Order Hybrid Recurrent Neural Network with Intrinsic Physics

This paper presents a reduced-order framework integrating singular value decomposition (SVD) with a hybrid gated recurrent unit (GRU)–long short-term memory (LSTM) network to predict long-term xenon dynamics in nuclear reactor cores. Traditional methods often rely on precise initial xenon and iodine distributions, involve high complexity, and incur significant computational costs. While dynamic mode decomposition performs well for short-term forecasts, its utility diminishes over extended horizons.

To address this, our approach first applies SVD to project reactor data into a low-dimensional latent space that intrinsically preserves governing physics. A hybrid GRU-LSTM architecture combining GRU efficiency in modeling short-term dependencies and LSTM robustness in capturing long-term memory is then trained within this reduced-order space to predict spatiotemporal xenon dynamics. The framework eliminates reliance on explicit physics equations while maintaining physical consistency.

Numerical experiments on the Hualong Pressurised Reactor 1000 (HPR1000) demonstrated improved accuracy in the xenon distribution predictions and the power peak factor estimation compared to traditional methods. Robustness tests under varying noise levels confirmed the model’s stability and scalability, highlighting its potential to enhance reactor safety and operational efficiency.